Recommendations tuned to your workload’s own behavior — not the average.
Before
After
$0/mo
unlocked across the fleet
0
workloads optimized
0
OOM-kills introduced
How it knows
Most tools pick a percentile and call it done. We ask three questions of every workload — and combine the answers into one safe recommendation.
For container worker
Baseline
01What does this workload normally need?
0.420 cores
Growth
02Is it trending up, down, or steady?
+5% headroom
Burstiness
03How spiky is its day-to-day?
+18% headroom
Recommended
0.000cores
Was
1.000
–48%
The diff · what you ship
The YAML on one tab. Seven days of evidence on the other.
Per-container YAML. Side-by-side. Copy either side.
Pick your aggregation. Avg through Max. 7, 14, or 30 days.
Dismiss what doesn’t fit. One click. We respect the cooldown.
- name: worker
resources:
requests:
cpu: 1000m
memory: 2Gi
limits:
cpu: 2000m
memory: 4Gi- name: worker
resources:
requests:
cpu: 320m # P95 × 1.05 × 1.18
memory: 780Mi # P99 × 1.04 × 1.22
limits:
cpu: 480m # request × 1.5
memory: 1.0Gi # max-observed × 1.30line 4
1000m 320m
-680m
line 5
2Gi 780Mi
-1.2Gi
line 7
2000m 480m
-1.52
line 8
4Gi 1.0Gi
-3.0Gi
Trend
Steady
Volatility (CV)
0.18 · Low
OOM kills
0 in 7d
Restarts
0 in 7d
Why it’s safe · four guardrails
Four checks. Every recommendation.
OOM-Aware
Out-of-memory kills your pod. So our recommended memory limit sits above the highest point we’ve seen — never an average that pretends rare spikes don’t happen.
Memory · 7 days
covers the peak
Two real spikes. An average misses the second. Our recommendation doesn’t.
After a crash
The kernel caps memory at the limit before the crash, so what we observed is lower than what was really needed. After kills, our recommendation pushes memory higher until the cycle stops.
7d · 3 kills detected
headroom restored
Kills cap what we can observe. Our recommendation pushes the limit higher until the cycle stops.
Bursty CPU
When peak usage runs far past normal load, we recommend removing the CPU limit so the spike can stretch instead of getting throttled. No tail-latency surprises in production.
Bursty CPU · 7 days
recommended ceiling: none
Big spikes go up. They don’t hit a wall.
Replica-aware
A 10% saving on one pod isn’t the story — the dollar number on the recommendation is per-pod savings multiplied by your replica count. HPA-scaled workloads included.
Replica-aware savings
$35.70 / pod × 40 pods
Total
$1428/ mo
Per-pod savings scaled by your replica count — including HPA workloads. The dollar number is the workload total, not a single pod.
The workflow · you stay in control
Read-only agent. Same review path as everything else.
Step 01
One click. The snippet is on your clipboard.
Current · Recommended
Step 02
Paste into whatever you already use.
Helm · Kustomize · Terraform · ArgoCD · Flux
Step 03
Your CI, your review, your rollout. We re-measure after 7 days.
7-day cooldown · auto re-analysis
Connect your first cluster in 5 minutes. The first 100 vCPU is on us, forever. Above it, the rate tiers from $1.99 down to $0.60 — same features, every plan.